diff --git "a/7tAzT4oBgHgl3EQf-f6Q/content/tmp_files/load_file.txt" "b/7tAzT4oBgHgl3EQf-f6Q/content/tmp_files/load_file.txt" new file mode 100644--- /dev/null +++ "b/7tAzT4oBgHgl3EQf-f6Q/content/tmp_files/load_file.txt" @@ -0,0 +1,527 @@ +filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf,len=526 +page_content='Topic Segmentation Model Focusing on Local Context Jeonghwan Lee, Jiyeong Han, Sunghoon Baek, Min Song* Yonsei University jeonghwan.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='ai@yonsei.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='kr, jiyoung181@yonsei.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='kr, sunghoon.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='baek@yonsei.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='kr, min.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='song@yonsei.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='ac.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='kr Abstract Topic segmentation is important in understanding scientific documents since it can not only provide better readability but also facilitate downstream tasks such as information retrieval and question answering by creating appropriate sections or paragraphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In the topic segmentation task, topic coherence is critical in predicting segmentation boundaries.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Most of the existing models have tried to exploit as many contexts as possible to extract useful topic-related information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, additional context does not always bring promising results, because the local context between sentences becomes inco- herent despite more sentences being supplemented.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To allevi- ate this issue, we propose siamese sentence embedding lay- ers which process two input sentences independently to get appropriate amount of information without being hampered by excessive information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Also, we adopt multi-task learn- ing techniques including Same Topic Prediction (STP), Topic Classification (TC) and Next Sentence Prediction (NSP).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' When these three classification layers are combined in a multi-task manner, they can make up for each other’s limi- tations, improving performance in all three tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We exper- iment different combinations of the three layers and report how each layer affects other layers in the same combination as well as the overall segmentation performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The model we proposed achieves the state-of-the-art result in the Wiki- Section dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Introduction Nowadays, we can easily access vast amounts of scientific documents such as PubMed and Wikipedia.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' A lot of re- searchers are studying ways to effectively use these docu- ments in areas like information retrieval (IR), question an- swering (QA) and search engine.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, applying previ- ous IR models (or QA models) directly on these documents is impossible because most of them assume an input size of at most a paragraph while these documents consist of multi- ple paragraphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Furthermore, extracting crucial parts of each document does not necessarily require the whole document to be used.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' For example, to search for similar papers on a topic of interest we can simplify the problem by calculating cosine similarity between sections of each document rather than full text to save resources.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Corresponding author.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Copyright © 2023, Association for the Advancement of Artificial Intelligence (www.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='aaai.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='org).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' All rights reserved.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Which topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Which topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Which topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Which topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Which topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Sentence 1 Sentence 2 Sentence 3 Sentence n Sentence n-1 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Same topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' & Consecutive?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Same topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' & Consecutive?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Same topic?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' & Consecutive?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Figure 1: A window with a size of 1 slides through the entire sentence, predicting the topic of each sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' At the same time, the model determines whether the two sentences are in the same topic and whether the two sentences are consecu- tive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' These are where topic segmentation can be used.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Topic segmentation divides a document into segments with respect to the topic coherence of each segment.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' A well-divided doc- ument according to the topics provides better readability, making it easier for the readers to find the desired infor- mation in the document.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Most importantly, it can facilitate downstream-tasks such as IR and QA.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Although most of the existing topic segmentation models take topic coherence into consideration when dividing a doc- ument, they don’t undergo the process of classifying topic labels for each sentence, even when these topic labels are useful for inferring topic coherence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Most importantly, in- spired by Neural Text Segmentation Model (Koshorek et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2018), these models are designed to take block of text as in- put, which possibly hinders understanding local context of the input text (Xing et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We try to tackle the above issues by adopting a siamese network to encode two input sentences independently and putting them through a multi-task learning algorithm that in- cludes topic classification and other auxiliary tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' First, in order to deal with two input sentences independently, we construct our model in a siamese network with sen- tence embeddings from a Sentence Transformer (Reimers arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='01935v1 [cs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='CL] 5 Jan 2023 and Gurevych 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This method allows our model to pre- serve local context between the two input sentences without being overwhelmed by excessive information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We consider topic segmentation as a Same Topic Predic- tion (STP) between two input sentences, following Aumiller et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, because our model processes only one sentence at a time to preserve its unique information, the model cannot observe context information across sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To alleviate this issue, we add two auxiliary tasks to cap- ture local context information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' One of them is Topic Classi- fication(TC) which predicts the exact topic of the input sen- tences through a topic classification layer to assist STP with a detailed topic information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The other is a Next Sentence Prediction (NSP) layer (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019), which supports the model in understanding the relationship between con- secutive sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Figure 1 simply shows how our model works.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To sum up, our model deals with two input sentences in- dependently via the siamese sentence embedding layer that preserves local context of input sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Also, we show that connecting tasks that utilize same input sentences to ex- tract different features in the sentence in a multi-task manner improves topic segmentation performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Consequently, our model achieves state-of-the-art in the topic segmentation task using the WikiSection dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Related Work Topic Segmentation Koshorek et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2018) solved topic segmentation task as a supervised neural network model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Block of text that consists of several sentences is fed into the model and the model pre- dicts whether each sentence should be a segmentation point.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Badjatiya et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2018) introduced k-sized left and right supporting sentences, where neighboring k number of sen- tences support injecting context into input sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' How- ever, Xing et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2020) pointed out that ”local context” was more important than ”global context” in topic segmentation task, implying that excessive context might decrease the per- formance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The information from various sentences can hin- der predicting label of a single sentence due to deterioration in the model’s understanding of local context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2019) proposed Sector which includes a topic embedding layer in their architecture.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This topic em- bedding layer is implemented for topic classification and the result of topic classification is then used for segmentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Aumiller et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2021) treated topic segmentation as a Same Topic Prediction(STP) between two input paragraphs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' STP determines whether two input paragraphs refer to the same topic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' They also experimented diverse sampling meth- ods, and among these methods we adopt consecutive sam- pling.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Details about consecutive sampling will be explained at section .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Sentence Embedding Sentence embedding is a method of capturing the seman- tic relationships among words in a sentence (Conneau et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2017).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Quality of sentence embedding is critical especially in the topic segmentation task, because the task inevitably has to capture as much information as possible from long sentences as well as short ones.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Koshorek et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2018) uti- lized Bi-LSTM to generate sentence embedding where word embedding vectors are extracted from Word2Vec with each word in a sentence as input, fed into the Bi-LSTM layer one by one, and the final sequence representation was made by max-pooling over the output of the LSTM.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' After the introduction of BERT, Reimers and Gurevych (2019) proposed Sentence BERT specialized in creating sentence embeddings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Sentence BERT is a fine-tuned ver- sion of BERT trained on NLI(Natural Language Inference) and STS(Semantic Textual Similarity) task.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To handle input sentences effectively, the authors adopted siamese network which encodes each sentence independently and concate- nates the encoded sentences to be fed into a classification layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Consequently, Sentence BERT has better capability of dealing with long sentences, because the model understands high-level context of these sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' As another branch of Sentence BERT, SimCSE was pro- posed (Gao, Yao, and Chen 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The authors applied con- trastive learning to forming sentence embeddings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' They tried unsupervised method and showed that dropout could work as data augmentation and this prevented representation col- lapse.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' They also empirically and theoretically proved that contrastive learning objective was suitable for regularizing anisotropic space of a language model’s embedding to be more uniform and it aligned positive pairs better in a super- vised setting as a result.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lukasik et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2020) proposed Cross-segment BERT.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' They used pre-trained BERT in which left and right context were separated via [SEP] token and encoded the sequence of word-piece tokens into sentence representations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Aumiller et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2021) used Sentence BERT for a sentence encoder which is known to have substantial capability of understand- ing high-level context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Proposed Approach Architecture Our model follows the typical architecture of text segmen- tation models: a sentence embedding layer followed by a segment classifier, which is replaced by a Same Topic Pre- diction layer in our model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, our model takes two input sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To handle them independently, our model composes sentence embed- ding layer in siamese network form, so that the model re- ceives an appropriate amount of information to predict the label.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The encoded sentences are then fed into the topic classification layer one by one.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' By passing each sentence through the layer, the model acquires topic-related informa- tion of the sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Also, we adopt NSP layer to capture semantic relationship between the two sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Finally, STP layer predicts whether the sentences belong to the same topic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We have k documents D1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='Dk that D consist of n num- ber sentences s1, .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=', sn, and the sentences are paired con- secutively;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' [(s1, s2), (s2, s3), .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='..' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=', (sn−2, sn−1), (sn−1, sn)].' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Each si(i ≤ n) is assigned a topic label ti which describes topic label of ith sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Models STP Loss TC Loss NSP Loss STP+TC 4 1 STP+NSP 1 1 STP+TC+NSP 4 1 4 Table 1: Designated loss weights for each layer in case of multi-task learning The sentence embedding layer encodes each input sen- tences si and si+1 and the encoded sentences are repre- sented as u and v, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Figure 2 shows the overview of our model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Siamese Sentence Embedding Layers from Sentence Transformer We propose siamese sentence embedding layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In our model, Sentence Transformer encodes each sen- tence from two input sentences independently at the entry level.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Then, the encoded sentences are concatenated before being fed into the STP layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This method aims to preserve each sentence’s unique information while acquiring local context between the two sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Multi Task Learning Our model has a total of three clas- sification layers and we train them in a multi-task manner.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Topic classification layer: Topic classification layer is designed to capture exact topic information of a sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Topic classification is a multi-class classification that pre- dicts the topic of an input sentence out of 30 labels for en city and 27 labels for en disease dataset (Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019) .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This layer takes u and v one by one and predicts each topic label ti and ti+1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' NSP layer: NSP layer is fed with u;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' v;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' |u − v| (Reimers and Gurevych 2019) and the layer predicts NSP label.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This layer aims to supplement STP layer’s limitation where STP layer can only determine whether the two input sentences are in the same topic and cannot determine if the sentences are actually consecutive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' By adding NSP, the model can capture the semantic relationship between the two sentences, so the model can figure out whether the sentences are consecutive.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' NSP layer must go with consecutive sampling which will be explained below.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' STP layer: STP layer is provided with u;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' v;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' |u − v| again and finally predicts segmentation label that is used to draw segmentation points in places where the two sentences be- long to different topics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Multi task learning: When the three layers are combined in a multi-task manner, they can make up each other’s limi- tations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Since STP is based on binary classification, its task is much simpler than Topic Classification that is based on multi-class classification.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, since STP cannot cap- ture the exact topic label of input sentences, Topic classi- fication provides this information to the STP layer to help determine segmentation boundaries.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The effect of NSP is ex- plained above.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Loss weight: Because the losses from each layer are all different, there is a need to adjust the weights among the losses for improved model performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We decide the weights for each loss by running numerous manual experi- ments and calculate the total loss using a weighted sum of en city en disease Docs 19,539 3,590 Topics 30 27 Table 2: The number of documents and topics for en city and en disease the three losses from each classification layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Table 1 sum- marizes how each loss is weighted.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Consecutive Sampling To make the model more robust, we add negative samples to the dataset by adopting consecutive sampling (Aumiller et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In consecutive sampling, all samples come from the same document.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We have a document Da and a sentence sti i ∈ Da where the superscript ti refers to topic label of si.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We pick one positive sample and two negative samples.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The positive sam- ple sti+1=ti i+1 ∈ Da is consecutive to si and the first nega- tive sample stk=ti k̸=i+1 is from the same topic as si, but not con- secutive to si .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Finally, the second negative sample stl̸=ti l is from different topics, which is naturally considered not consecutive to si .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Experiment Dataset We use WikiSection (Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019) for training and evaluating our model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' WikiSection covers two distinct do- mains: city and disease.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Each domain has 19,539 and 3,590 documents, respectively, with various topics in each docu- ment.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In total, there are 30 and 27 topics for each domain.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The dataset is divided into 70% training, 10% validation and 20% test sets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Table 2 gives statistics of the dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Experimental Setup We use nltk sentence tokenizer1 to split the documents into sentence units and apply consecutive sampling only on the training dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Table 3 gives the data statistics after ap- plying sentence split and consecutive sampling.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We imple- ment all-MiniLM-L12-v2 and from Sentence-Transformers2 for our sentence encoder.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We set the maximum epoch size to 14 but save the model only when the validation Pk scores best.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Batch size is 48, learning rate is 1e−6 and LinearLR scheduler is applied with the default parameters setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Metric For a comprehensive evaluation, we used Pk, WindowDiff and micro F1 score to evaluate our models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We use Pk score for making comparisons between our models and all other baseline models, and WindowDiff is used to evaluate ours and Cross-segment BERT that we implemented.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' F1 score is used for the purpose of ablation study on our models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 1NLTK :: Natural Language Toolkit 2Pretrained Models — Sentence-Transformers documentation (sbert.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='net) 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# Pooling Pooling 𝑢 𝑣 Softmax Classifier (NSP) 𝑢;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑣;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' |𝑢 − 𝑣| 0 or 1 Sentence Transformer Sentence Transformer 𝑡!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑡!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# Softmax Classifier (Topic Classification) Softmax Classifier (STP) 0 or 1 Figure 2: The overview of our model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Two input sentences which are considered consecutive are fed into Sentence Transformer independently and encodes each input sentence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Each max-pooled encoded sentence, represented by u and v respectively, is fed into Topic classification layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Before being fed into NSP layer and STP layer, we make concatenated feature u;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' v;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' |u−v|.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Using u;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' v;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' |u − v|, NSP layer predicts if the two sentences are consecutive and STP layer finally determines whether they belong to same topic.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 79.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='5 95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='4 95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='7 73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='7 86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='5 40 50 60 70 80 90 100 TC only STP only STP+TC STP+NSP STP+TC+NSP F1 Micro F1 Scores of STP, TC and NSP (en_city) STP TC NSP 54.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='6 88 88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='4 45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='9 46.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 64 64.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='5 40 50 60 70 80 90 100 TC only STP only STP+TC STP+NSP STP+TC+NSP F1 Micro F1 Scores of STP, TC and NSP (en_disease) STP TC NSP Figure 3: Figure of F1 scores with combination of different task layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In case of TC-only model, STP output is 1 if the results of topic classification on each sentence refer to the same topic otherwise 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' en city en disease Train 1,690,103 336,459 Valid 85,072 16,285 Test 168,924 31,110 Table 3: The number of rows after applying nltk sentence tokenizer and consecutive sampling.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Consecutive sampling is applied only on the trainset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Pk Pk (Beeferman, Berger, and Lafferty 1999) is a proba- bility that a segmentation model performs an incorrect seg- mentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' While a sliding window of size k passing over the sentences, the status (0 or 1) is determined by whether the two ends of the window are in the same segment or in different segments.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Pk is calculated by counting unmatched cases between the ground truths and predicted values.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' As in many previous studies, we set the window size k to half the average segment length of the ground truths.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Dataset en city en disease Metric Pk WinDiff Pk WinDiff SEC>T+emb 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='5 26.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 Transformer2 BERT 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 18.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 BiLSTM + BERT 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 28.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 Cross-segment BERT n context = 2 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='4 27.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='4 33.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='9 59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 Cross-segment BERT n context = 4 18.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 34.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='3 Cross-segment BERT n context = 6 45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 50.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 34.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 57.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 TC-only 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 17.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 41.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='5 45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='4 STP-only 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 STP + TC 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='7 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='0 STP + NSP 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='9 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='6 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 STP + TC + NSP 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='6 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='7 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='7 Table 4: Test Pk and WindowDiff scores of baseline models and our models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Note that the WinDiff metric is used only in our models and the Cross-segment BERT models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We reimplement Cross-segment BERT ourselves following their official codes.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' WindowDiff WindowDiff (Pevzner and Hearst 2002) is an improved metric from Pk in that it alleviates the impact of false negative penalty and segment size distribution.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Similar to Pk, WindowDiff score also uses sliding win- dow and compares the ground truths with the predicted val- ues.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, this metric also takes the number of bound- aries into consideration.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' It is closer to the ground truth when the models get a lower score in both Pk and WindowDiff.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Baseline Models We compare our model with competitive neural text seg- mentation baselines 1) SEC>T+emb (Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019), 2) Transformer2 BERT (Lo et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2021), a framework based on two transformers, where one is a pre-trained transformer for encoding sentences and the other is a transformer for seg- mentation, 3) Bi-LSTM + BERT (Xing et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020), that is based on a hierarchical attention Bi-LSTM network, and 4) Cross-segment BERT (Lukasik et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020), which handles left and right context simultaneously using a BERT encoder.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We adopt the results of SEC>T+emb from Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2019), Transformer2 BERT and Bi-LSTM + BERT from Xing et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We implement Cross-segment BERT ourselves following their official code while apply- ing diverse size of context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Results and Analysis We report evaluation results on Figure 3 and Table 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Fig- ure 3 summarizes how combination of each classification layer affects their F1 scores.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Table 4 shows performance comparison between our model and other baseline models in Pk and WindowDiff, respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Our proposed mod- els, except for TC-only model, outperform all the baseline models by a large margin.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Effect of MTL Figure 3 shows F1 scores derived from combinations of tasks mentioned above.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We can see that MTL is effective in improving the performance, which ap- plies to not only the performance of STP that is responsible segmentation but also the performances of TC and NSP.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' This is believed to be because, as we pointed out in the section , the layers make up for each other’s limitations by extract- ing different features from same input sentences that assist understanding semantic information.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' STP-only vs TC-only In order to verify the effectiveness of STP layer, we also experiment TC-only model, which is close to Sector in that segmentation is performed only using topic labels.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Pk and WindowDiff of TC-only model are much higher than those of STP-only model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Poor classification performance of Topic classification directly causes this phe- nomenon.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Figure 3 indicates that F1 scores of TC-only model are significantly lower than those of STP-only.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Be- cause topic classification layer is based on multi-class classi- fication, which is more difficult than the binary classification of STP-only.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' STP vs NSP Although STP and NSP have the same ar- chitecture, STP’s F1 scores are always higher than NSP’s in both datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We assume that this difference is derived from the difference in the information that STP and NSP focus on.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' STP-only determines whether the two sentences belong to the same topic, so it only pays attention to topic differences between two sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In other words, due to the nature of the task, STP does not consider the relationship between two input sentences.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' However, in the NSP task, the layer faces difficulties as two sentences may not be consecutive even if they belong to the same topic because of our consecutive sampling.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Thus, NSP must find the semantic relationship be- tween the two sentences as well as topic coherence, which makes the task tricky.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' How the number of contexts affects the performance To show the importance of local context, we implement Cross-segment BERT (Lukasik et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020) by applying di- verse size of context on the model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Table 4 shows that raising context size rather deteriorates the performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We conjec- ture that this is because the more sentences there are, the more likely for different topics to be mingled, which likely interferes the model from understanding local context with Left context ����!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='"\' n_context = 1 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='"\' n_context = 4 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='"\' n_context = 2 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$# 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='$& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "% 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' "& 𝑠!' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='"\' n_context = 3 Right context Left context Right context Left context Right context Left context Right context Figure 4: Effect of context size on prediction.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The color of the box represents the topic of the sentence and the red line represents supporting context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The vertical dotted line represents a segmentation point between si and si+1 while the dotted box describes that the two sentences are not segmented.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' si and si+1 should be divided, since they belong to different topics, but are not segmented in cases of n context = 3 and n context = 4 .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' overflow of noise.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Because processing multiple sentences si- multaneously using a left and right context structure rather adds noise to the contexts, we choose to encode the two in- put sentences independently.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Figure 4 explains this local context capturing error.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' The models in Figure 4 are all expected to create a segmentation point between si and si+1 , but models with larger con- text sizes fail to split the two sentences, because segmenta- tion only takes into account the overall context of each side.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' As the context size increases, the model suffers from gener- alization and interprets left and right contexts as similar even when the two specific sentences refer to different topics and hence should be segmented.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Also, Cross-segment BERT encodes left and right context simultaneously.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Because Cross-encoder inevitably makes context of one input sentence influence the other (Humeau et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019), the unique information of each context can change unexpectedly.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Therefore, we process each sentence independently via siamese sentence embedding in order to preserve the original local context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Dealing with Scientific Documents We can also find that the scores for en disease are underperforming compared to that for en city.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We assume that this result due to the fact that en disease is more science domain specific (i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' biol- ogy) while en city covers relatively general topics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Arnold et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' (2019) commented that documents in en disease are described in a precise language, but on the other hand those in en city are described in a common language.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Consider- ing that our backbone, miniLM was pre-trained on general documents like Wikipedia, the result seems natural.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' To improve the performance on en disease, we implement SPECTER (Cohan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020) which was trained on scien- tific papers using Sci-BERT as the backbone model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' As shown in table 5, Pk and WinDiff improved by 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='6 and 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='8 respectively in en disease compared to the miniLM based model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We attribute the improvement to SPECTER’s understanding of scientific documents.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We expect the scores Pk WinDiff en city 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='6 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='2 en disease 12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='1 12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='9 Table 5: Test Pk and WindowDiff scores of SPECTER based Topic Segmentation Model with STP+TC+NSP for en disease to be improved if we use more biology do- main specific model like BioBERT as the backbone.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Interestingly, although the number of parameters of SPECTER was twice as large as that of miniLM (i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 768 vs 384), there was no improvement in the performance in en city.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' From this result, we can again confirm that domain knowledge is critical to the performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Conclusion and Future Work In this work, we propose our topic segmentation model which consists of siamese sentence embedding layer from Sentence Transformer and three classification layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' With several different experiments, we show that our proposed model outperforms all the existing models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' We also find that combining Same Topic Prediction, Topic Classification and Next Sentence Prediction in a multi-task manner increases segmentation performance.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Moreover, we empirically show the importance of local context in topic segmentation task.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Contrary to the popu- lar belief, increasing the number of context can rather de- grade the performance due to generalization of local context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Our experiment indicates that narrowing context through our siamese sentence embedding layer can be effective in pre- serving local context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Future work can highlight on the theoretical approach to local context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Although we empirically showed the influence of context size to the model performance in this paper, we did not concentrate on how we can determine which input sentences can provide substantial information in performing segmentation tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' If we can infer each sentence’s signifi- cance in prediction, we expect the model to capture the im- portant sentences autonomously, consequently making the model agnostic to the context size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Acknowledgments This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020- 0-01361, Artificial Intelligence Graduate School Program (Yonsei University)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' References Arnold, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Schneider, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Cudr´e-Mauroux, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Gers, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and L¨oser, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' SECTOR: A Neural Model for Coherent Topic Segmentation and Classification.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Aumiller, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Almasian, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lackner, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Gertz.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Structural text segmentation of legal documents.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceed- ings of the Eighteenth International Conference on Artificial Intelligence and Law, 2–11.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computing Ma- chinery.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Badjatiya, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Kurisinkel, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Gupta, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Varma, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Attention-based Neural Text Segmentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Beeferman, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Berger, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Lafferty, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 1999.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Statistical Models for Text Segmentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Machine Learning, 34: 177– 210.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Cohan, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Feldman, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Beltagy, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Downey, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Weld, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' SPECTER: Document-level Representation Learn- ing using Citation-informed Transformers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics, 2270–2282.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Online: Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Conneau, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Kiela, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Schwenk, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Barrault, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Bor- des, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Supervised Learning of Universal Sentence Representations from Natural Language Inference Data.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 670–680.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Devlin, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Chang, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='-W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lee, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Toutanova, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 2019 Con- ference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies, volume 1, 4171–4186.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Gao, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Yao, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Chen, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' SimCSE: Simple Con- trastive Learning of Sentence Embeddings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 6894–6910.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Online and Punta Cana, Dominican Republic: Association for Computational Lin- guistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Humeau, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Shuster, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lachaux, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content='-A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' ;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Weston, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Poly-encoders: Transformer Architectures and Pre- training Strategies for Fast and Accurate Multi-sentence Scoring.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Koshorek, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Cohen, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Mor, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Rotman, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Berant, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Text Segmentation as a Supervised Learning Task.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 2018 Conference of the North Ameri- can Chapter of the Association for Computational Linguis- tics: Human Language Technologies, volume 2, 469–473.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lo, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Jin, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Tan, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Liu, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Du, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Buntine, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Transformer over Pre-trained Transformer for Neu- ral Text Segmentation with Enhanced Topic Coherence.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Findings of the Association for Computational Linguistics: EMNLP 2021, 3334–3340.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Lukasik, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Dadachev, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Papineni, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Sim˜oes, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Text Segmentation by Cross Segment Attention.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Pro- ceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 4707–4716.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Pevzner, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Hearst, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2002.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' A critique and im- provement of an evaluation metric for text segmentation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Computational Linguistics, 28(1): 19–36.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Reimers, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Gurevych, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 2019 Conference on Empirical Meth- ods in Natural Language Processing and the 9th Interna- tional Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 3982–3992.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computa- tional Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Xing, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Hackinen, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Carenini, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=';' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' and Trebbi, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Improving Context Modeling in Neural Topic Segmenta- tion.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Lan- guage Processing, 626–636.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'} +page_content=' Association for Computational Linguistics.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/7tAzT4oBgHgl3EQf-f6Q/content/2301.01935v1.pdf'}